Academics

  /  

Courses

  /  

Course Descriptions

MECH_ENG 495: (BME_468)Decision-making in the wild: Measurement and models

Description

Decision-making in the wild: Measurement and models

The class will be a survey of the most recent science and quantitative models of high stakes non-reactive decision making, focused primarily on non-human mammals and birds. There will be a significant computational component, in Python. Modern artificial intelligence/reinforcement learning (RL) methods do not perform well in high-stakes scenarios where very few data samples may be available. For example, when a “predator” is introduced into a task that an RL agent is learning, state-of-the-art algorithms seek to learn the source of reward prediction error, which in this case means many “lethal” interactions. Yet animals, when a novel predator is introduced, quickly learn through observation, and change tactics. This is one of many sources of an immense learning efficiency gap, and is part of the reason the best RL agents, such as AlphaGo, require five to six orders of magnitude more energy than human go players for similar task performance. The class will look at the evolutionary basis of high efficiency rapid learning and decision making in animals, current behavioral data, current understanding of mechanisms from the neuroscience and psychology literature, and algorithms to model the phenomena in the low data limit that animals excel in.